cklxx/webchain-elements

WebChain Elements — human browser trajectories as element-level decision steps

10,930 steps (2,555 tasks, 243 real websites) from WebChain (webagentlab, CC-BY-4.0) turned into one-step decision examples in the text format browser-agent decision heads use: the task goal, the page (URL, title, visible text), the list of interactive elements, the actions taken so far, and the gold operation + gold element the human chose.

split	steps	tasks	websites	CLICK / TYPE_TEXT / PRESS_ENTER
train	8,435	1,935	199	7,273 / 1,053 / 109
test	2,495	620	44	2,109 / 364 / 22

The test split holds out whole websites (none of its 44 sites appears in train). Tasks are mostly multi-constraint (filters, sorts, dates) on shopping, travel, real-estate and job sites.

Fields

goal         the task, e.g. "Look for the lowest-priced women's double-breasted trench coats."
website      the site's primary host; domain / intent: WebChain's labels
url, title   the page the step was taken on; page_text: its visible text (<= 6,000 chars)
candidates   interactive elements (≤ 60, page order): {id, kind: click|fill, label, role, value?, checked?, ...}
history      earlier steps of the same trajectory: {action, kind, text, page_changed}
gold_op      CLICK | TYPE_TEXT | PRESS_ENTER
gold_id      the chosen candidate's id (click:N / fill:N; "press_enter" for PRESS_ENTER)
gold_label   its label; gold_text: the typed text for TYPE_TEXT

How it was built

WebChain records, per step, a CSS selector plus links to the page's accessibility tree (AX) and DOM snapshot. For 3,000 randomly chosen trajectories (seed 0), every step's AX tree was fetched and flattened; the gold element was recovered exactly by resolving the step's CSS selector in the DOM snapshot and following its data-imean-axt-id to the AX node (text matching only as a fallback — it disagreed with the selector on ~13 % of a sample, so it is never the default). Candidates are the page's interactive AX nodes plus some non-semantic text nodes (the clicked target is often a <span>/<div>), capped at 60 in page order. Steps whose gold element has no label or whose label is duplicated among the candidates were dropped (they cannot be learned from text), as were hover / drag / copy / paste steps. WebChain has no scroll and no explicit "task done" steps, so neither appears here.

Converter: code/finetune/convert_webchain.py in cklxx/laya-browser, whose model v19s was trained on the train split (held-out test: click operation 0.94, click target top-1 0.48).

License and attribution

CC-BY-4.0, derived from WebChain by webagentlab (paper, dataset). Page texts come from public websites as recorded in WebChain.